1 /* 2 * IPV4 GSO/GRO offload support 3 * Linux INET implementation 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU General Public License 7 * as published by the Free Software Foundation; either version 8 * 2 of the License, or (at your option) any later version. 9 * 10 * UDPv4 GSO support 11 */ 12 13 #include <linux/skbuff.h> 14 #include <net/udp.h> 15 #include <net/protocol.h> 16 17 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb, 18 netdev_features_t features, 19 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb, 20 netdev_features_t features), 21 __be16 new_protocol, bool is_ipv6) 22 { 23 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb); 24 bool remcsum, need_csum, offload_csum, gso_partial; 25 struct sk_buff *segs = ERR_PTR(-EINVAL); 26 struct udphdr *uh = udp_hdr(skb); 27 u16 mac_offset = skb->mac_header; 28 __be16 protocol = skb->protocol; 29 u16 mac_len = skb->mac_len; 30 int udp_offset, outer_hlen; 31 __wsum partial; 32 bool need_ipsec; 33 34 if (unlikely(!pskb_may_pull(skb, tnl_hlen))) 35 goto out; 36 37 /* Adjust partial header checksum to negate old length. 38 * We cannot rely on the value contained in uh->len as it is 39 * possible that the actual value exceeds the boundaries of the 40 * 16 bit length field due to the header being added outside of an 41 * IP or IPv6 frame that was already limited to 64K - 1. 42 */ 43 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL) 44 partial = (__force __wsum)uh->len; 45 else 46 partial = (__force __wsum)htonl(skb->len); 47 partial = csum_sub(csum_unfold(uh->check), partial); 48 49 /* setup inner skb. */ 50 skb->encapsulation = 0; 51 SKB_GSO_CB(skb)->encap_level = 0; 52 __skb_pull(skb, tnl_hlen); 53 skb_reset_mac_header(skb); 54 skb_set_network_header(skb, skb_inner_network_offset(skb)); 55 skb->mac_len = skb_inner_network_offset(skb); 56 skb->protocol = new_protocol; 57 58 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM); 59 skb->encap_hdr_csum = need_csum; 60 61 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM); 62 skb->remcsum_offload = remcsum; 63 64 need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb)); 65 /* Try to offload checksum if possible */ 66 offload_csum = !!(need_csum && 67 !need_ipsec && 68 (skb->dev->features & 69 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) : 70 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM)))); 71 72 features &= skb->dev->hw_enc_features; 73 74 /* The only checksum offload we care about from here on out is the 75 * outer one so strip the existing checksum feature flags and 76 * instead set the flag based on our outer checksum offload value. 77 */ 78 if (remcsum) { 79 features &= ~NETIF_F_CSUM_MASK; 80 if (!need_csum || offload_csum) 81 features |= NETIF_F_HW_CSUM; 82 } 83 84 /* segment inner packet. */ 85 segs = gso_inner_segment(skb, features); 86 if (IS_ERR_OR_NULL(segs)) { 87 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset, 88 mac_len); 89 goto out; 90 } 91 92 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL); 93 94 outer_hlen = skb_tnl_header_len(skb); 95 udp_offset = outer_hlen - tnl_hlen; 96 skb = segs; 97 do { 98 unsigned int len; 99 100 if (remcsum) 101 skb->ip_summed = CHECKSUM_NONE; 102 103 /* Set up inner headers if we are offloading inner checksum */ 104 if (skb->ip_summed == CHECKSUM_PARTIAL) { 105 skb_reset_inner_headers(skb); 106 skb->encapsulation = 1; 107 } 108 109 skb->mac_len = mac_len; 110 skb->protocol = protocol; 111 112 __skb_push(skb, outer_hlen); 113 skb_reset_mac_header(skb); 114 skb_set_network_header(skb, mac_len); 115 skb_set_transport_header(skb, udp_offset); 116 len = skb->len - udp_offset; 117 uh = udp_hdr(skb); 118 119 /* If we are only performing partial GSO the inner header 120 * will be using a length value equal to only one MSS sized 121 * segment instead of the entire frame. 122 */ 123 if (gso_partial && skb_is_gso(skb)) { 124 uh->len = htons(skb_shinfo(skb)->gso_size + 125 SKB_GSO_CB(skb)->data_offset + 126 skb->head - (unsigned char *)uh); 127 } else { 128 uh->len = htons(len); 129 } 130 131 if (!need_csum) 132 continue; 133 134 uh->check = ~csum_fold(csum_add(partial, 135 (__force __wsum)htonl(len))); 136 137 if (skb->encapsulation || !offload_csum) { 138 uh->check = gso_make_checksum(skb, ~uh->check); 139 if (uh->check == 0) 140 uh->check = CSUM_MANGLED_0; 141 } else { 142 skb->ip_summed = CHECKSUM_PARTIAL; 143 skb->csum_start = skb_transport_header(skb) - skb->head; 144 skb->csum_offset = offsetof(struct udphdr, check); 145 } 146 } while ((skb = skb->next)); 147 out: 148 return segs; 149 } 150 151 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb, 152 netdev_features_t features, 153 bool is_ipv6) 154 { 155 __be16 protocol = skb->protocol; 156 const struct net_offload **offloads; 157 const struct net_offload *ops; 158 struct sk_buff *segs = ERR_PTR(-EINVAL); 159 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb, 160 netdev_features_t features); 161 162 rcu_read_lock(); 163 164 switch (skb->inner_protocol_type) { 165 case ENCAP_TYPE_ETHER: 166 protocol = skb->inner_protocol; 167 gso_inner_segment = skb_mac_gso_segment; 168 break; 169 case ENCAP_TYPE_IPPROTO: 170 offloads = is_ipv6 ? inet6_offloads : inet_offloads; 171 ops = rcu_dereference(offloads[skb->inner_ipproto]); 172 if (!ops || !ops->callbacks.gso_segment) 173 goto out_unlock; 174 gso_inner_segment = ops->callbacks.gso_segment; 175 break; 176 default: 177 goto out_unlock; 178 } 179 180 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment, 181 protocol, is_ipv6); 182 183 out_unlock: 184 rcu_read_unlock(); 185 186 return segs; 187 } 188 EXPORT_SYMBOL(skb_udp_tunnel_segment); 189 190 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb, 191 netdev_features_t features, 192 unsigned int mss, __sum16 check) 193 { 194 struct sock *sk = gso_skb->sk; 195 unsigned int sum_truesize = 0; 196 struct sk_buff *segs, *seg; 197 unsigned int hdrlen; 198 struct udphdr *uh; 199 200 if (gso_skb->len <= sizeof(*uh) + mss) 201 return ERR_PTR(-EINVAL); 202 203 hdrlen = gso_skb->data - skb_mac_header(gso_skb); 204 skb_pull(gso_skb, sizeof(*uh)); 205 206 /* clear destructor to avoid skb_segment assigning it to tail */ 207 WARN_ON_ONCE(gso_skb->destructor != sock_wfree); 208 gso_skb->destructor = NULL; 209 210 segs = skb_segment(gso_skb, features); 211 if (unlikely(IS_ERR_OR_NULL(segs))) { 212 gso_skb->destructor = sock_wfree; 213 return segs; 214 } 215 216 for (seg = segs; seg; seg = seg->next) { 217 uh = udp_hdr(seg); 218 uh->len = htons(seg->len - hdrlen); 219 uh->check = check; 220 221 /* last packet can be partial gso_size */ 222 if (!seg->next) 223 csum_replace2(&uh->check, htons(mss), 224 htons(seg->len - hdrlen - sizeof(*uh))); 225 226 seg->destructor = sock_wfree; 227 seg->sk = sk; 228 sum_truesize += seg->truesize; 229 } 230 231 refcount_add(sum_truesize - gso_skb->truesize, &sk->sk_wmem_alloc); 232 233 return segs; 234 } 235 EXPORT_SYMBOL_GPL(__udp_gso_segment); 236 237 static struct sk_buff *__udp4_gso_segment(struct sk_buff *gso_skb, 238 netdev_features_t features) 239 { 240 const struct iphdr *iph = ip_hdr(gso_skb); 241 unsigned int mss = skb_shinfo(gso_skb)->gso_size; 242 243 if (!can_checksum_protocol(features, htons(ETH_P_IP))) 244 return ERR_PTR(-EIO); 245 246 return __udp_gso_segment(gso_skb, features, mss, 247 udp_v4_check(sizeof(struct udphdr) + mss, 248 iph->saddr, iph->daddr, 0)); 249 } 250 EXPORT_SYMBOL_GPL(__udp4_gso_segment); 251 252 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb, 253 netdev_features_t features) 254 { 255 struct sk_buff *segs = ERR_PTR(-EINVAL); 256 unsigned int mss; 257 __wsum csum; 258 struct udphdr *uh; 259 struct iphdr *iph; 260 261 if (skb->encapsulation && 262 (skb_shinfo(skb)->gso_type & 263 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) { 264 segs = skb_udp_tunnel_segment(skb, features, false); 265 goto out; 266 } 267 268 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4))) 269 goto out; 270 271 if (!pskb_may_pull(skb, sizeof(struct udphdr))) 272 goto out; 273 274 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) 275 return __udp4_gso_segment(skb, features); 276 277 mss = skb_shinfo(skb)->gso_size; 278 if (unlikely(skb->len <= mss)) 279 goto out; 280 281 /* Do software UFO. Complete and fill in the UDP checksum as 282 * HW cannot do checksum of UDP packets sent as multiple 283 * IP fragments. 284 */ 285 286 uh = udp_hdr(skb); 287 iph = ip_hdr(skb); 288 289 uh->check = 0; 290 csum = skb_checksum(skb, 0, skb->len, 0); 291 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum); 292 if (uh->check == 0) 293 uh->check = CSUM_MANGLED_0; 294 295 skb->ip_summed = CHECKSUM_UNNECESSARY; 296 297 /* If there is no outer header we can fake a checksum offload 298 * due to the fact that we have already done the checksum in 299 * software prior to segmenting the frame. 300 */ 301 if (!skb->encap_hdr_csum) 302 features |= NETIF_F_HW_CSUM; 303 304 /* Fragment the skb. IP headers of the fragments are updated in 305 * inet_gso_segment() 306 */ 307 segs = skb_segment(skb, features); 308 out: 309 return segs; 310 } 311 312 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb, 313 struct udphdr *uh, udp_lookup_t lookup) 314 { 315 struct sk_buff *p, **pp = NULL; 316 struct udphdr *uh2; 317 unsigned int off = skb_gro_offset(skb); 318 int flush = 1; 319 struct sock *sk; 320 321 if (NAPI_GRO_CB(skb)->encap_mark || 322 (skb->ip_summed != CHECKSUM_PARTIAL && 323 NAPI_GRO_CB(skb)->csum_cnt == 0 && 324 !NAPI_GRO_CB(skb)->csum_valid)) 325 goto out; 326 327 /* mark that this skb passed once through the tunnel gro layer */ 328 NAPI_GRO_CB(skb)->encap_mark = 1; 329 330 rcu_read_lock(); 331 sk = (*lookup)(skb, uh->source, uh->dest); 332 333 if (sk && udp_sk(sk)->gro_receive) 334 goto unflush; 335 goto out_unlock; 336 337 unflush: 338 flush = 0; 339 340 for (p = *head; p; p = p->next) { 341 if (!NAPI_GRO_CB(p)->same_flow) 342 continue; 343 344 uh2 = (struct udphdr *)(p->data + off); 345 346 /* Match ports and either checksums are either both zero 347 * or nonzero. 348 */ 349 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) || 350 (!uh->check ^ !uh2->check)) { 351 NAPI_GRO_CB(p)->same_flow = 0; 352 continue; 353 } 354 } 355 356 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */ 357 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr)); 358 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb); 359 360 out_unlock: 361 rcu_read_unlock(); 362 out: 363 NAPI_GRO_CB(skb)->flush |= flush; 364 return pp; 365 } 366 EXPORT_SYMBOL(udp_gro_receive); 367 368 static struct sk_buff **udp4_gro_receive(struct sk_buff **head, 369 struct sk_buff *skb) 370 { 371 struct udphdr *uh = udp_gro_udphdr(skb); 372 373 if (unlikely(!uh)) 374 goto flush; 375 376 /* Don't bother verifying checksum if we're going to flush anyway. */ 377 if (NAPI_GRO_CB(skb)->flush) 378 goto skip; 379 380 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check, 381 inet_gro_compute_pseudo)) 382 goto flush; 383 else if (uh->check) 384 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check, 385 inet_gro_compute_pseudo); 386 skip: 387 NAPI_GRO_CB(skb)->is_ipv6 = 0; 388 return udp_gro_receive(head, skb, uh, udp4_lib_lookup_skb); 389 390 flush: 391 NAPI_GRO_CB(skb)->flush = 1; 392 return NULL; 393 } 394 395 int udp_gro_complete(struct sk_buff *skb, int nhoff, 396 udp_lookup_t lookup) 397 { 398 __be16 newlen = htons(skb->len - nhoff); 399 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff); 400 int err = -ENOSYS; 401 struct sock *sk; 402 403 uh->len = newlen; 404 405 /* Set encapsulation before calling into inner gro_complete() functions 406 * to make them set up the inner offsets. 407 */ 408 skb->encapsulation = 1; 409 410 rcu_read_lock(); 411 sk = (*lookup)(skb, uh->source, uh->dest); 412 if (sk && udp_sk(sk)->gro_complete) 413 err = udp_sk(sk)->gro_complete(sk, skb, 414 nhoff + sizeof(struct udphdr)); 415 rcu_read_unlock(); 416 417 if (skb->remcsum_offload) 418 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM; 419 420 return err; 421 } 422 EXPORT_SYMBOL(udp_gro_complete); 423 424 static int udp4_gro_complete(struct sk_buff *skb, int nhoff) 425 { 426 const struct iphdr *iph = ip_hdr(skb); 427 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff); 428 429 if (uh->check) { 430 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM; 431 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr, 432 iph->daddr, 0); 433 } else { 434 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL; 435 } 436 437 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb); 438 } 439 440 static const struct net_offload udpv4_offload = { 441 .callbacks = { 442 .gso_segment = udp4_ufo_fragment, 443 .gro_receive = udp4_gro_receive, 444 .gro_complete = udp4_gro_complete, 445 }, 446 }; 447 448 int __init udpv4_offload_init(void) 449 { 450 return inet_add_offload(&udpv4_offload, IPPROTO_UDP); 451 } 452